feat: P1.7 data simulation agent for small-sample augmentation

[Mola-maker]

Inserts between P1.5 (data cleaning) and P2 (modeling) to augment cleaned CSVs whose row count falls below MIN_ROWS_FOR_MODELING (30).

Augmentation method

Gaussian-perturbation bootstrap:

• bootstrap-sample rows from the cleaned dataframe
• for each numeric column, add N(0, sigma * col_std) noise (sigma = 0.05 relative)
• non-numeric columns are copied verbatim
• tag each row with _sim_origin column ("real" | "simulated")
• per-column Kolmogorov-Smirnov stat is logged as quality signal

Integration

• agents/data_simulation.py: new DataSimulationAgent (pure stats, no LLM; scipy optional with numpy fallback)
• main.py: p1_7 closure + PhaseSpec(on_error="skip", record_experience) inserted between P1.5 and P2; --start choices extended
• agents/experience_recorder.py: _PHASE_PROMPTS["P1.7"] prompt + _extract_phase_context branch + _PHASE_NAME entry
• ui/server.py: PHASE_META, PHASE_ORDER, PHASE_COMPLETE_MAP updated

Safety

• Never overwrites cleaned_.csv — writes augmented_.csv alongside
• on_error="skip": missing data / read failures don't halt pipeline
• _sim_origin column name safe against existing fabrication regex
• ctx.data_simulation.simulated_files lists absolute output paths for any future validator whitelist

Smoke-tested with 3 synthetic CSVs (20-row → augment, 100-row → skip sufficient, text-only → skip no-numeric). All KS stats < 0.1.